Hebbrix MCP Server

A Model Context Protocol server that gives any AI agent persistent memory, a temporal knowledge graph, and outcome-based learning, backed by Hebbrix.

This repository is the MIT-licensed MCP adapter. Durable storage, retrieval, graph processing, reasoning, and Outcome Memory are provided by the Hebbrix cloud API; this repository does not claim that the backend is self-hostable or open source.

Your agent forgets everything when the session ends. This fixes that, and goes further than a plain memory store:

• Memory — store, search, correct, and version facts across sessions
• Knowledge graph — entities, relationships, timelines, and "what was true at time X"
• Reasoning — ask how confident the agent should be before acting, and log outcomes so it improves
• Outcome Memory — learn which action works for each customer and context from delayed results, with safe baselines and inspectable uncertainty

Works with Claude Desktop, Claude Code, Cursor, Cline, Continue, and any other MCP client.

Release compatibility

Version 0.5.10 preserves upstream indexing receipts across writes, updates, and polling. Graph status distinguishes related memories from entity relationships. The companion API's claim-grounding-v28 contract verifies named-subject evidence consistently across cold and cached retrieval and preserves purpose and direction; the adapter never bypasses upstream grounding to manufacture an answer.

Version 0.5.11 adds response-local diagnostics (request ID, build, grounding and safety versions, and search cache path). An ask fallback includes separate reasoning and retrieval diagnostics. No credentials or memory content are included in diagnostics. synthesis_status distinguishes synthesized, retrieval_only, and abstained. failure_category distinguishes unknown facts, unverified candidates, unavailable verification, service failures, quota exhaustion, and malformed evidence receipts. Legitimate abstention is not a transport error. These labels do not weaken the evidence gate.

Reproduce the full customer set and additional distractor checks with python scripts/verify_recall.py --require-graph against a local API first; use --mcp http://localhost:8099/mcp for HTTP. Remote fixture writes require --allow-remote. The verifier creates isolated guests, tests both result limits before and after asynchronous processing, then deletes its collections. This targeted regression suite is not a representative accuracy benchmark; conservative abstentions and probabilistic synthesis remain possible.

Version 0.5.9 makes structured failures real MCP errors over both stdio and hosted HTTP, preserves authoritative recall when GraphRAG abstains by returning an explicitly retrieval-only result, and adds bounded graph-enrichment status polling. Version 0.5.8 added the complete tenant-scoped procedure lifecycle (including idempotent deletion), and preserves authoritative batch readiness receipts. It also retains 0.5.7's API-owned grounding and abstention envelope: missing, malformed, degraded, or ungrounded receipts fail closed with empty evidence.

The hosted server and PyPI package expose their exact package version during MCP initialization. The API exposes its immutable deployment build through X-Hebbrix-Build and GET /v1/health/build; OpenAPI info.version identifies the stable HTTP contract rather than a mutable deployment.

Fastest setup: hosted, no account

For an HTTP-capable MCP client, this is the entire setup:

{ "mcpServers": { "hebbrix": { "url": "https://mcp.hebbrix.com/mcp" } } }

On the first MCP handshake, Hebbrix creates an isolated free guest memory and keeps its credential in a Secure, HttpOnly session cookie. There is no signup, email, dashboard, local process, or API key to paste. A compatible MCP HTTP client automatically sends that cookie on later requests. Add your own bearer key at any time if you want to use an existing Hebbrix account instead.

Quick start (no account needed)

Add this to your MCP client config. On first run with no API key, the server mints a free agent account automatically (no email, no dashboard, ~2-4 seconds via proof-of-work) and saves it to ~/.hebbrix/config.json.

{
  "mcpServers": {
    "hebbrix": { "command": "uvx", "args": ["hebbrix-mcp"] }
  }
}

Restart the client. Done — your agent now has persistent memory.

The free agent account includes 300 learning events and 2,000 retrievals, and expires 14 days after last use if unclaimed. The first tool result, material quota/status changes, and every constrained-state result carry a hebbrix_usage block; hebbrix_account_status returns it on demand at any time.

Keep it forever (same key, all memories carry over, unlocks the free monthly tier):

uvx hebbrix-mcp claim --email you@example.com

Claude Code plugin (recommended)

Install as a Claude Code plugin and Claude starts every session already knowing you — a SessionStart hook auto-loads your compiled Hebbrix profile into context, and the memory tools are wired up in one step:

/plugin marketplace add Hebbrix/hebbrix-mcp
/plugin install hebbrix@hebbrix

That's it. No account needed (agent mode mints one on first run); set your api_key in the plugin config to use your own account instead. The hook degrades gracefully — a brand-new profile just shows (none yet) until you've saved a few facts, and it never blocks a session.

Configuration

Get an API key at hebbrix.com/dashboard/api-keys to use your own account instead of agent mode.

{
  "mcpServers": {
    "hebbrix": {
      "command": "uvx",
      "args": ["hebbrix-mcp"],
      "env": {
        "HEBBRIX_API_KEY": "mem_sk_...",
        "HEBBRIX_COLLECTION_ID": "your-default-collection-uuid"
      }
    }
  }
}

claude mcp add hebbrix -- uvx hebbrix-mcp

{
  "mcpServers": {
    "hebbrix": { "command": "uvx", "args": ["hebbrix-mcp"] }
  }
}

The env var always wins over saved agent-mode credentials.

Environment variables

All optional. With nothing set, the server starts in agent mode.

Available Tools

A server-level instruction block teaches the model when to reach for each tool, so a well-behaved agent searches before answering and remembers what matters without being told.

Memory

• hebbrix_remember - Store a fact, decision, or preference.
  • content (string, required): the memory text
  • tags (list, optional), collection_id (string, optional)
  • extract (bool, default false): false stores the text exactly (one memory); true starts a tracked fact-extraction job and may create several atomic memories
  • wait_for_extraction (bool, default true): for smart ingestion, poll for up to 20 seconds and return normalized atomic memories. Set false for immediate acknowledgement, then call hebbrix_extraction_status with the returned job id.
  • wait_for_index (bool, default true): requests a bounded wait for memory-search availability. Check the returned searchable acknowledgement: if false, the write is durable but indexing is not confirmed; poll hebbrix_get(id) and do not repeat the write. It does not cover asynchronous knowledge-graph enrichment; use hebbrix_graph_status for durable completion. Its related_memory_count counts neighboring memories, not entity relationships; zero neighbors does not mean zero extracted edges.
• hebbrix_extraction_status - Poll a smart-ingestion job until its created/updated memories or terminal error are available.
• hebbrix_remember_many - Store many facts in one call (one round-trip, one rate-limit hit). Pass facts (list of strings). Falls back to sequential writes on free/agent tiers.
• hebbrix_search - Semantic search (hybrid vector + BM25 + graph retrieval).
  • query (string, required), limit (int, optional), collection_id (string, optional)
  • min_score (float, default 0.0): drop weak matches — zero-relevance padding is always dropped; raise this to filter noise so you don't pay tokens for it.
• hebbrix_get - Fetch one memory by id, with metadata.
• hebbrix_update - Correct a memory in place (old versions are kept).
• hebbrix_forget - Delete a memory by id.
• hebbrix_list - List recent memories.
• hebbrix_history - See how a memory changed over time.
• hebbrix_mark_used - Reinforce a memory you actually used (helpful=True strengthens it, False weakens it) so recall improves over time.
• hebbrix_export - Export a whole collection (memories + graph entities + profile) as JSON or Markdown, in one call.
• hebbrix_import - The inverse of export: import a list of facts, an export JSON, or notes/markdown into a collection (restore a backup, migrate, or seed from CLAUDE.md).

Knowledge graph — Hebbrix automatically extracts entities and relationships from the memories you write, on every tier including agent mode, so all the graph reads below (entities, timelines, traversal, contradictions) work in agent mode too. Only explicit graph write / inference operations require a Pro plan.

• hebbrix_search_entities - List known entities (people, orgs, tools, places).
• hebbrix_entity_timeline - What was true about an entity, and when.
• hebbrix_graph_query - Traverse relationships out from a named entity; pass a timestamp for point-in-time truth. Results are trimmed (from/to/type/valid_from), not raw backend payloads. (Free-text questions: use hebbrix_ask.)
• hebbrix_graph_status - Check one memory's asynchronous graph readiness, optionally polling for up to 30 seconds. Readiness comes from durable, source-revision-bound delivery acknowledgements—not elapsed time, search readiness, or an empty graph query. It distinguishes successful enrichment with no related facts from a still-processing, failed, or temporarily unreadable graph.
• hebbrix_contradictions - Surface facts that conflict with each other.

Procedural memory

• hebbrix_create_procedure - Store a scoped condition/action procedure.
• hebbrix_list_procedures / hebbrix_get_procedure - Inspect owned procedures.
• hebbrix_update_procedure - Update mutable fields; ownership scope is immutable.
• hebbrix_execute_procedure - Execute a procedure and record the execution.
• hebbrix_delete_procedure - Idempotently delete a procedure and its executions. The API returns the same 204 for deleted, absent, and foreign-tenant IDs so the tool cannot reveal another tenant's identifier.

October 2 follow-up: the setup/report tools, automatic v2 selection, decision proposal priors and structured paraphrase matching below require the matching October 2 outcome-followup backend. Inspect /v1/release before use. No learning-performance superiority, reliable model compliance with feedback, or automatic execution is established. These tools never grant permission.

Reasoning & account

• hebbrix_ask - One-call GraphRAG. Ask a natural-language question; it searches memory, synthesizes an answer with an LLM, and cites the memory ids it used. If synthesis abstains while authoritative search has grounded evidence, it returns that evidence as synthesis_status: "retrieval_only" instead of silently losing recall or pretending synthesis succeeded.
• hebbrix_confidence - Read separate memory support and action-confidence signals, with ASK/REVIEW/ACT advice. Pass policy/action/context for exact configured evidence. Memory similarity alone never earns ACT. Advice is not permission.
• hebbrix_log_decision - Record a decision and its outcome; feeds future confidence. Right after a hebbrix_confidence check you can log just the outcome — the description auto-fills from what you asked.
• hebbrix_choose_action - Choose among repeatable strategies and create a decision receipt before acting. Supports per-user/context policies and bounded exploration; neither a receipt nor caller-supplied outcomes establish causality or execution authority. adaptive_exploration=true requires an explicitly configured posterior-sampling policy and low-risk exploration opt-ins. The new default auto enables sampling only for v2 low-risk opted-in actions; old or unconfigured policies stay conservative. prior_action and prior_strength supply a bounded, decaying proposal for one server-selected decision, never a reported outcome or a permission grant.
• hebbrix_setup_policy - Atomically enroll context and configure a new policy. Requires owner-declared action risk and exploration permission; cannot reset an existing policy. Unknown/high-risk actions remain conservative.
• hebbrix_learning_report - Read a bounded scoped weekly report of observed leaders and recorded evidence-readiness transitions. Current adaptive-policy uplift is not identified; descriptive action rates are not causal lift.
• hebbrix_context_schema - Enroll the immutable declared context before the first decision. Fields use {name: {values: [...], required: true}}. Optional backoff never drops required fields. Only change contracts on owner instruction.
• hebbrix_configure_policy - Revision-checked configuration of strategy, bounded priors, decay, window and action risk. Requires explicit owner instruction; never silently lower risk or replace the action target to obtain ACT advice.
• hebbrix_policy_advice - Read a scoped evidence card with ranked actions, uncertainty, reported failure attempts, descriptive change alerts and an exploration candidate. Not a chosen-action receipt or permission to execute.
• hebbrix_report_outcome - Close that decision loop later with success, a bounded reward, or configured business metrics. Corrections replace prior evidence instead of double-counting it.
• hebbrix_learning_insights - Inspect posterior probabilities, credible intervals, effective evidence, and optional chronological-holdout policy readiness checks for one customer policy.
• hebbrix_list_collections - List the memory spaces this key can use.
• hebbrix_account_status - Tier, usage, limits, and expiry.
• hebbrix_claim_start / hebbrix_claim_verify - Optionally keep an accountless guest memory permanently, without changing its collection or losing data.

Every tool publishes explicit MCP safety annotations. hebbrix_claim_start is marked as an external side effect because it sends email; deletion is marked destructive; reads are marked read-only. The six-digit claim code is declared as a write-only password field and is never logged or returned by this server. MCP hosts still control their own tool-call history, so configure the host to redact secret inputs if it persists conversation or tracing data.

The server also exposes a hebbrix://profile resource and a context prompt that inject the user's compiled profile.

Make Hebbrix the agent's memory

The server ships an instruction block telling the model to use Hebbrix for anything it would "remember." But some hosts (notably Claude Code) have their own file-based memory whose instructions live at the system-prompt level and can outrank an MCP server's instructions — so the agent may quietly write notes to a local file instead of Hebbrix.

The reliable fix is one line in your project's CLAUDE.md (or your assistant's system prompt / rules file):

## Memory
Use the Hebbrix MCP server as the single source of truth for long-term memory.
When you would remember, note, or save anything durable, call `hebbrix_remember`
(and `hebbrix_search` to recall). Do not write memory to local files or the
host's built-in memory.

Cursor users: add the same to .cursorrules. This puts the preference at the level the host respects, so Hebbrix wins consistently.

Running modes

Local (default) — stdio. What the quick start does: one process per client.

Self-hosted HTTP — one instance, your machines:

HEBBRIX_API_KEY=mem_sk_... uvx hebbrix-mcp --transport streamable-http
# serves http://127.0.0.1:8080/mcp

Hosted — nothing to run and no account required. Point any HTTP-capable MCP client at the official hosted endpoint. The first handshake creates an isolated guest memory and a Secure, HttpOnly session cookie automatically:

{ "mcpServers": { "hebbrix": {
  "url": "https://mcp.hebbrix.com/mcp"
}}}

To use an existing Hebbrix account instead, add its API key (get one at hebbrix.com/dashboard/api-keys):

{ "mcpServers": { "hebbrix": {
  "url": "https://mcp.hebbrix.com/mcp",
  "headers": { "Authorization": "Bearer mem_sk_..." }
}}}

Self-hosted multi-tenant — one instance, many users. Same shape on your own infra. By default every request authenticates with its own Authorization header:

HEBBRIX_MCP_MULTI_TENANT=1 HEBBRIX_MCP_HOST=0.0.0.0 uvx hebbrix-mcp --transport streamable-http

Or run the container (multi-tenant by default, GET /healthz for load-balancer probes):

docker build -t hebbrix-mcp . && docker run -p 8080:8080 hebbrix-mcp

In multi-tenant mode, the server resolves each authenticated key's default collection automatically. An explicit collection_id still overrides it.

How it works

┌──────────────────┐   MCP (stdio or HTTP)   ┌─────────────┐    HTTPS     ┌──────────┐
│ Claude / Cursor / │ ───────────────────────→│ hebbrix-mcp │─────────────→│ Hebbrix  │
│ Cline / any agent │      tool calls         │   (this)    │   REST API   │  cloud   │
└──────────────────┘                          └─────────────┘              └──────────┘

This package owns no durable memory state. Tool calls become REST calls against your Hebbrix tenant; memories, embeddings, the knowledge graph, and retrieval all live in the Hebbrix backend. The hosted accountless path keeps only a signed identity cookie in the MCP client so multiple stateless replicas can serve it. Delete the local package and your backend memories are still there.

Agent-mode accounts never break mid-task: when a limit is reached you get a structured error with a resolve field, not a failure. Writes stop before reads; reads keep working; the account goes read-only before it expires.

Debugging

Inspect the server with the MCP Inspector:

npx @modelcontextprotocol/inspector uvx hebbrix-mcp

Common issues:

• HTTP 401 on every call — the key is wrong or revoked. Unset HEBBRIX_API_KEY, delete ~/.hebbrix/config.json, and restart to re-provision, or paste a fresh key from the dashboard.
• Agent mode won't start (auto-signup unavailable) — signup may be at daily capacity or your network blocks the API. Set HEBBRIX_API_KEY instead.
• claim says EMAIL_IN_USE — claiming needs an email with no existing Hebbrix account. Use a fresh address (a you+agent@gmail.com alias works).
• A memory isn't searchable immediately — wait_for_index=true requests a bounded wait, not an unconditional guarantee. Check searchable; if false, poll hebbrix_get(id) and preserve the receipt. Do not repeat an accepted write.
• A just-written fact's entities aren't in the graph yet — knowledge-graph enrichment (entities, timelines, graph queries) runs asynchronously after the write and is not covered by wait_for_index. It typically lands within ~30s; the write response's graph_enrichment: "processing" signals it's still in flight.

Development

git clone https://github.com/Hebbrix/hebbrix-mcp
cd hebbrix-mcp
python -m pip install uv==0.8.4
uv sync --frozen --extra dev
uv run --frozen pytest tests/ -q -W error  # offline; no network or key needed
hebbrix-mcp                 # starts in agent mode on stdio

See CONTRIBUTING.md and CHANGELOG.md.

An opt-in integration check reproduces conversational recall, corrections, fresh-process retrieval, abstention, isolation, and disposable collection cleanup:

uv run --frozen python scripts/verify_recall.py --api http://localhost:8000

It creates synthetic data in two new guest collections and deletes them in a cleanup block. Remote targets require --allow-remote; --mcp URL selects the hosted transport and --require-graph checks asynchronous graph enrichment too.

License

MIT — see LICENSE.

Links

• Hebbrix documentation
• MCP integration guide
• Model Context Protocol